Anthropic’s new AI model, Claude Mythos Preview, has become one of the most closely watched—and most unsettling—developments in artificial intelligence this year. The company has not released it to the general public. Instead, it has placed the model inside Project Glasswing, a controlled initiative meant to give select organizations access for defensive cybersecurity work. Anthropic says Mythos is a general-purpose, unreleased frontier model whose coding abilities show that AI can now surpass all but the most skilled humans at finding and exploiting software vulnerabilities.
That sentence alone explains why the question has become so urgent. Is Mythos too dangerous for ordinary users?
The responsible answer is yes—at least for now. The issue is that Mythos appears to sit at a dangerous threshold. The model is powerful enough to find software flaws, chain them together, and potentially turn them into working exploits. Anthropic’s own transparency materials say Mythos demonstrated a “notable leap” in cyber capabilities, including the ability, after an initial user prompt, to autonomously discover and exploit zero-day vulnerabilities in major operating systems and web browsers.
For ordinary consumers, that is not a productivity tool. It is a capability with public-safety implications.
What Mythos Is Supposed to Do
Anthropic’s public framing is defensive. Project Glasswing is designed to help major technology companies and security organizations identify and repair vulnerabilities before attackers can exploit them. The company says Claude Mythos Preview is being made available only to a limited set of partners for defensive cybersecurity purposes, with access controls and safeguards.
The company’s developer documentation also says Mythos is offered separately as a research preview for defensive cybersecurity workflows, with invitation-only access and no self-serve sign-up.
Anthropic is not presenting Mythos as a chatbot for the public. It is not being marketed as an ordinary writing assistant, search companion, coding helper, or office tool. It is being held back precisely because the company believes the model’s cyber abilities are too powerful to distribute broadly without preparation and safeguards.
But the same fact that makes Mythos useful for defense also makes it dangerous in the wrong hands. A tool that can find hidden weaknesses in software can help defenders patch them. The same tool can also help attackers exploit them.
The Unauthorized Access Problem
The public concern grew sharper after reports that unauthorized users gained access to Mythos. Reuters reported on April 21 that a small group of unauthorized users accessed Anthropic’s Mythos model. According to the report, the users gained access on the same day Anthropic announced plans to release the model to a limited number of companies for testing. Anthropic said it was investigating a report of unauthorized access through a third-party vendor environment.
The Guardian reported similar details, noting that Anthropic confirmed it was investigating the claim and that the model had not been released publicly because of its ability to enable cyberattacks. The report said the users were not believed to have run cybersecurity prompts, but the incident still raised alarms about how easily powerful systems can be kept out of the wrong hands.
That is the heart of the concern. If Mythos is too risky to release publicly, then the security around Mythos must be stronger than ordinary software access controls. A model this powerful cannot be treated like a beta product that accidentally leaks through a vendor portal. If the danger is real, then the containment must be real too.
Why Experts Are Worried
The fear surrounding Mythos is found in Anthropic’s own technical writing which says, non-experts can use Mythos Preview to find and exploit sophisticated vulnerabilities. The company said engineers without formal security training have asked Mythos to find remote code execution vulnerabilities overnight and awakened to a complete working exploit. In other cases, researchers developed scaffolding that allowed Mythos to turn vulnerabilities into exploits without human intervention.
That is a dramatic change in who can do serious cyber work. Traditionally, high-level exploitation required deep expertise, time, and patience. Mythos appears to compress that process. It does not simply explain code. It can help move from discovery to exploitation. That lowers the barrier between curiosity and harm.
The Guardian also reported that the UK’s AI Security Institute described Mythos as a “step up” from previous models and said it could carry out attacks requiring multiple actions and discover weaknesses in IT systems without human intervention. The model reportedly completed a 32-step cyberattack simulation in three out of ten attempts.
Even if three out of ten sounds unimpressive, it is still alarming. In cybersecurity, attackers do not need every attempt to succeed. They just need one path in. If AI systems can repeatedly test paths, refine attempts, and chain weaknesses together, the speed of attack changes.
The Moral Question Beneath the Technical One
For Christians, the Mythos debate is not only about software security. It is about wisdom, stewardship, and the moral limits of human power.
Scripture does not oppose tools. Human beings are creative because we are made in the image of a Creator. We build, cultivate, organize, discover, and design. But Scripture is also honest about what happens when human ability outruns human humility. The tower of Babel was not condemned because people used technology. It was judged because human ambition became detached from reverence, restraint, and accountability.
That is the spiritual warning behind powerful AI. We are not simply asking, “Can we build it?” We are asking, “Who governs it?” “Who benefits?” “Who is harmed if it fails?” “What happens when the people who created the tool can no longer fully control its use?”
Mythos forces those questions into the open.
If a model can help defenders secure the internet, then there is a moral case for using it carefully. Hospitals, banks, power systems, public agencies, churches, schools, and ordinary families all depend on software. Hidden vulnerabilities can endanger real people. A tool that finds those weaknesses before criminals do could be an act of protection.
But if that same model can accelerate cyberattacks, then broad public release would be reckless. Power requires discipline. Capability requires accountability. Innovation without moral restraint can quickly become danger disguised as progress.
Too Dangerous for Users—or Too Dangerous Without Governance?
The better question may not be whether Mythos is too dangerous for users. It may be whether Mythos is too dangerous for the current AI governance system.
Anthropic says it is keeping Mythos limited and testing cyber safeguards on less capable models first. In announcing Claude Opus 4.7, the company said Mythos would remain limited while Anthropic tests new safeguards on models with less advanced cyber capabilities.
That is a prudent step. But the unauthorized-access reports show that good intentions are not enough. Safety depends on vendor security, auditing, legal accountability, misuse monitoring, human oversight, and clear consequences for failure.
What Christians Should Watch
Christians need to be morally awake.
Mythos represents a turning point because it shows how quickly AI is moving from “answering questions” to performing complex, high-stakes tasks. The question is no longer whether AI can write emails or summarize articles. The question is whether AI can act inside systems that shape public life, economic stability, national security, and human safety.
That should concern the church. But because Christians are called to test the spirit of the age, not simply consume it. We are called to ask whether our tools serve human flourishing, truth, justice, and neighbor-love—or whether they deepen fear, exploitation, and control.
At minimum, Mythos should not be available to ordinary users in its current form. It should remain restricted to vetted defensive use, with stronger safeguards, independent oversight, and transparent reporting when security incidents occur.
The deeper lesson is larger than Anthropic. We are entering an age where the most powerful tools may be too dangerous for mass release, too profitable to ignore, and too consequential to govern casually. That is precisely the kind of moment that requires wisdom.
The danger is not only that AI may become powerful.
The danger is that human beings may become powerful without becoming wise.
